AMap deploys a mixture-of-experts design that separates expert participation from compute cost

single source· 1 articles · confidence: medium · first seen 2026-09-17 20:00 UTC

What this means for you

If you serve a mixture-of-experts model under a latency budget, the idea is worth a look: the paper claims full expert participation at sparse compute cost, and it is deployed at scale, not just benchmarked. Code is public. No weights or API are offered, and no evaluation date is given for the benchmark gains.

AMap's recommendation system is running IntBMoE, a mixture-of-experts design (only part of the network computes each request) that separates three quantities existing designs tie together: how many experts feed a token's output, how many are computed, and how many expert-sized parameter sets must be stored. It holds the first at full while keeping the other two low, by composing experts from a small learned codebook of blocks and routing each token to a few. The paper reports deployment at hundreds of millions of users inside a 60ms latency budget and a 2.4% relative gain in UVCTR, an online click-through measure, in A/B testing. It gives no evaluation date for its benchmark comparisons.

Key facts

  • ·IntBMoE is deployed in AMap's generative recommendation system, serving hundreds of millions of users under a 60ms latency budget. source
  • ·The paper reports a 2.4% relative UVCTR gain in online A/B testing. source
  • ·The method pairs dense expert composition with sparse block execution, drawing blocks from a small learned codebook, one per entry. source
  • ·Code is available at https://github.com/AMAP-ML/DreamX-Rec/. source
  • ·The preprint is arXiv 2609.21346, posted 17 September 2026. source
  • ·Gains over sparse and dense MoE baselines are reported on image classification, with further experiments on language modelling and sequential recommendation, and no evaluation date is given. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire